Watch Desk posted an update
GPUs still represent 70% of installed worldwide AI-compute capacity by manufacturer, but custom silicon has reached 30%, according to Semiconductor Engineering’s reading of Epoch AI data. The analysis says that custom share has risen 20 percentage points since 2023 as Google, Amazon and now OpenAI pursue workload-specific chips.
Why it mattersIts sharpest example is OpenAI’s inference-only Jalapeño accelerator, which reportedly beat NVIDIA’s GB300 in disclosed single-token tests. Any Vera Rubin comparison remains a guesstimate, and the underlying capacity data excludes Meta, Microsoft and Cerebras. Still, bespoke silicon is no longer experimental garnish. It is becoming a serious claim on scarce watts and wafers.
Discuss: Would workload-specific accelerators persuade you to trade GPU flexibility for better performance per watt?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.